I'm guessing that the idea is to make it easy to enforce permissions at the application layer. You give permissions, and you get only cells that the current query-er is allowed to see. With HBase, it would be pretty easy to put permissions by the row (add a permission column, or column family if it's complicated enough), but if you want some columns in a row to have some permissions and some to have different ones, it would get unpleasant and inefficient fast.
And regardless, all of the filtering would have to occur at the application layer, meaning you'd have to wrap every get/scan to have it do the filtering for you. The Accumulo way also gets you some efficiency because it never even has to transfer the cells that get filtered by the permissions (or even fully read their content from disk, possibly).
Even though each cell isn't separately encrypted to get you true security at the cell level (which would destroy your performance, I'd guess), this seems like a huge win if you want to have permissions at the cell level.
And regardless, all of the filtering would have to occur at the application layer, meaning you'd have to wrap every get/scan to have it do the filtering for you. The Accumulo way also gets you some efficiency because it never even has to transfer the cells that get filtered by the permissions (or even fully read their content from disk, possibly).
Even though each cell isn't separately encrypted to get you true security at the cell level (which would destroy your performance, I'd guess), this seems like a huge win if you want to have permissions at the cell level.